sudo dnf install cuda-toolkit-12-4
export CUDA_HOME=/usr/local/cuda-12.4/
export TORCH_CUDA_ARCH_LIST="9.0a"
python setup.py develop
python test.py
python benchmark.py
Note: test.py is currently broken so that NVIDIA folks can repro the bug I'm seeing with cpfence. benchmark.py does not use the broken function. Please comment the last line of test.py if you just want to make pretty charts.
